Researchers from a Caltech lab say they have created some of the first AI that optimizes the gait of a lower body exoskeleton for the comfort of the person using it. Gait is the length and move of stride as a person walks. The way people walk can be pretty distinct. That’s why gait recognition is a form of biometric identification used in the wild today.

The gait optimization model comes from the Advanced Mechanical Bipedal Experimental Robotics Lab (AMBER) Lab at Caltech — more specifically, the Robotic Assisted Mobility Science (RoAMS) initiative that explores use of spinal cord stimulation, exoskeletons, and robotic prosthetics. Assistance for people with lower body impairments could help more than 6 million people in the United States and tens of millions around the world.

Exoskeletons have been considered for use in the workplace and by military, but they can also improve mobility and independence for people with physical disabilities. People who have lost use of their legs, have conditions like cerebral palsy, or use tools like walkers or standers may also find walking exoskeletons useful someday.

Previous methods, authors argue, use gait settings for robot on two feet, not people. “While existing methods can generate human-like walking gaits for bipedal robots, it is unlikely that these methods fulfill the preferences of individuals using robotic assistance,” the paper reads. “In the experiments, CoSpar consistently found user-preferred parameters of the exoskeleton’s walking gait, which suggests that it is a promising starting point for adapting and personalizing exoskeletons (or other assistive devices) to individual users.”

The name CoSpar comes from the fact that the algorithm applies coactive learning, an approach to defining interaction between a learning system and human user, to the dueling bandits technique for learning from preference feedback. Caltech researchers introduced the SELFSPARRING algorithm for solving dueling bandit problems in 2017.

The CoSpar algorithm was tested in simulations and in trials with able-bodied people at Caltech using Atalante, a lower body exoskeleton with 12 actuated joints being developed by the company Wandercraft.

The gait optimization algorithm was published by the IEEE International Conference on Robotics and Automation, aka the ICRA conference. Originally scheduled to take place in Paris, ICRA converted into a digital conference that started online on May 31.

The researchers believe CoSpar, which leverages preference-based learning (does the user like gait A or gait B?), is the first project to apply preference-based learning in conjunction with cocative learning to optimize dynamic crutchless walking based on human comfort. It both queries users to better understand an individual’s preferences and takes user suggestions to make improvements. Some pre-computed gaits are available so users can get started immediately. CoSpar updates this model with user feedback and uses it to select actions for new trials, elicit feedback, and “identify subregions of preferred gaits.”

“In the future, we plan to apply CoSpar toward optimizing over larger sets of gait parameters; this will likely require integrating the algorithm with techniques for learning over high-dimensional feature spaces. The method could also be extended beyond working with precomputed gait libraries to generating entirely new gaits or controller designs (e.g., via preference-based reinforcement learning),” the paper reads.

In other papers published at ICRA this week about robotics learning how to walk, ETH Zurich researchers also introduced DeepGait, reinforcement learning for four-legged robots to walk on non-flat terrain, climb stairs, or bridge unusually long gaps.